The DNA methylation profiles of AD and control samples from postmortem brain donors’ four brain regions: CRB, DLPFC, ERC, and HIPPO, were downloaded from Gene Expression Omnibus under accession number of GSE125895 (Semick et al., 2019). In CRB dataset, there were 24 AD and 43 control samples. In DLPFC dataset, there were 21 AD and 47 control samples. In ERC dataset, there were 20 AD and 49 control samples. In HIPPO dataset, there were 17 AD and 48 control samples. The DNA methylation levels were measured with Illumina HumanMethylation450 BeadChip and represented with normalized beta values processed with the R/Bioconductor package minfi¹. The processed data of 420,852 methylation probes in 269 samples were used for further analysis.

As mentioned in section “Data,” all samples in four datasets were represented by a large number of methylation features. Evidently, not all of these features give contributions for identification of AD samples. It is necessary to exclude irrelevant features first.

To quick complete this task, we employed the Boruta feature filtering (Kursa and Rudnicki, 2010). It is a powerful feature selection method based on RF. The Boruta method involves several steps. First, shuffled data from copies of the original data are created. Second, the RF is trained on the original and shuffled data to obtain feature importance. Third, the Z score for each feature is computed on the basis of this feature’s importance score. Fourth, a feature is tagged as important if and only if its Z score is greater than that of shadow features. Finally, these steps are repeated for all features. The features tagged as important are outputted as the outcome of this method, which are deemed to be relevant features.

In this study, we adopted the Boruta program obtained from https://github.com/scikit-learn-contrib/boruta_py, which was performed on each dataset with its default parameters. Selected features would be analyzed by the following mRMR method.

For the remaining methylation features, we further used mRMR method (Peng et al., 2005) to analyze them. This method calculates the mutual information (MI) between features and class labels for evaluating feature relevance and MI between features to assess the redundancy between features. Its original purpose was to find out a feature subset that had maximum relevance to class labels and minimum redundancies between features in this set. However, such problem is NP-hard. In view of this, it adopts a heuristic way. In this way, investigated features are sorted in a feature list, where important features are assigned high ranks. In the beginning, this list is empty. The method repeatedly selects one features from the remaining features such that it has maximum relevance to class labels and minimum redundancies to features already in the list. Such procedure stops until all features are in the list. This list is called mRMR feature list.

The present study used the mRMR program retrieved from http://penglab.janelia.org/proj/mRMR/. It was applied on each dataset, in which samples were represented by features selected by Boruta. Accordingly, one mRMR feature list was obtained on each dataset.

Although mRMR method deeply analyzed the importance of each methylation feature, it was still a problem which features can be selected. Thus, we employed the IFS method (Liu and Setiono, 1998) to determine the optimal number of features according to the integrating supervised classification algorithm. In brief, a series of feature subsets are extracted from the mRMR feature list yielded by the mRMR method at a step interval of 1. For instance, the first feature subset has the top ranked feature, and the second feature subset has the top 2 ranked features, and so on. For each feature subset, a classifier is trained based on the training data induced from this feature subset and the given classification algorithm. All classifiers are evaluated by 10-fold cross-validation (Kohavi, 1995). When all classifiers have been assessed, the best one can be discovered. This classifier is called the optimal classifier and the features used in such classifier are termed as the optimal features.

To execute the IFS method, one classification algorithm is necessary. In this study, we used the following three classification algorithms: RF (Breiman, 2001), SVM (Cortes and Vapnik, 1995), and PART (Frank and Witten, 1998).

In each dataset, control samples were about twice as many as AD samples. It is better to balance samples in these two classes. To this end, the widely used method, synthetic minority oversampling technique (SMOTE; Chawla et al., 2002) was adopted, which is an over-sampling method. It iteratively produces new samples for the minor class until the minor class has equal number of samples in major class. In detail, it first randomly selects a sample, say x, in the minor class. Then, a number of samples in this class are picked up such that these samples have smallest distance to x. The randomly selected sample, say y, from above samples and x are used to generate the new sample z, which is defined as the linear combination of x and y. The tool “SMOTE” in Weka was applied in this study. The SMOTE was only used in evaluation of performance of classifiers. The samples generated by it did not participate in evaluating the importance of methylation features.

Matthew correlation coefficients (MCC) is a widely used measurement for binary classification (Matthews, 1975; Zhao et al., 2019; Liang et al., 2020; Zhang et al., 2021b; Yang and Chen, 2022), which was adopted herein to assess the performance of the different classifiers within a 10-fold cross-validation (Kohavi, 1995; Zhao et al., 2019; Zhang et al., 2021a; Tang and Chen, 2022). The MCC can be computed as follows:

where TP, TN, FP, and FN indicate the sample number of true-positive, true-negative, false-positive, and false-negative predictions, respectively. MCC values range from -1 to +1 and achieves +1 when the classification model has the best performance.

In addition, we also calculated other five measurements: sensitivity (SN), specificity (SP), prediction accuracy (ACC), precision and F1-measure, to fully display the performance of different classifiers. They can be computed by

For each brain structure, one methylation dataset was collected. As a large number of methylation features were involved in each dataset, the Boruta method was first applied on each dataset to exclude irrelevant features. Important methylation features remained for each dataset, which are provided in Supplementary Table 1. The number of remaining features on each dataset is listed in Table 1. It can be observed that less than 200 features were kept on each dataset, greatly reducing the scope of investigated methylation features. These features were deemed to be essential for distinguishing AD patients from control samples on corresponding brain structure.

The features selected by Boruta method were further analyzed by the mRMR method according to Figure 1. One mRMR feature list was produced on each dataset, which is also provided in Supplementary Table 1. Such list would be investigated in the following analysis.

The mRMR method produced an mRMR feature list on each dataset, corresponding to one brain structure. Then, the IFS method was applied on such list, which integrated three classification algorithms.

For the CRB dataset, the mRMR feature list contained 124 methylation features. Accordingly, 124 feature subsets were generated in the IFS method by using step one. On each subset, a classifier was built with one of the three classification algorithms, which was further assessed by 10-fold cross-validation. The predicted results were counted as measurements listed in section “Performance Evaluation,” which are provided in Supplementary Table 2. To clearly display the performance of the classifiers based on one classification algorithm and different feature subsets, one IFS curve was plotted for each classification algorithm by setting MCC as Y-axis and number of used features as X-axis, as shown in Figure 2. It can be observed that the RF, SVM, and PART produced the highest MCC values of 0.968, 0.936, and 0.846, respectively. These values were obtained by using top 38, 16, and 6, respectively, features in the mRMR feature list. These features were the optimal ones for corresponding classification algorithm. Accordingly, the optimal RF/SVM/PART classifier can be built using corresponding optimal features. The ACC and F1-measure of these classifiers are listed in Table 2, whereas the other three measurements: SN, SP, and precision, are illustrated in Figure 3A. Evidently, the optimal RF classifier was best, followed by the optimal SVM classifier and optimal PART classifier.

For the other three datasets (ERC, DLPFC, and HIPPO datasets), the same IFS procedures were conducted. Detailed IFS results are provided in Supplementary Tables 3–5. Three IFS curves, corresponding to three classification algorithms, were plotted, as shown in Figures 4–6. From Figure 4 for ERC dataset, we can see that RF/SVM/PART yielded the highest MCC of 0.966/0.966/0.934, which was based on top 1/4/2 features in the list. Accordingly, we can build three optimal classifiers with corresponding features. Other measurements of these classifiers are provided in Table 2 and Figure 3B. Clearly, the optimal RF and SVM classifiers gave equal performance, whereas they were all superior to the optimal PART classifier. For the IFS results on DLPFC dataset, it can be observed from Figure 5 that three classification algorithms generated the highest MCC values of 0.967, 1.000, and 0.868, respectively. Top 6, 20, and 2, respectively, features in the list were adopted to produce these MCC values, which comprised the optimal features for the corresponding classification algorithm. With these optimal features, three optimal classifiers were built. Their detailed performance is listed in Table 2 and Figure 3C. The optimal SVM classifier gave the best performance, followed by the optimal RF and PART classifiers. As for the last dataset (HIPPO dataset), the highest MCC value for RF/SVM/PART was 0.926/1.000/0.920, which can be observed from Figure 6. This MCC was obtained by using top 15/18/34 features in the list. They were optimal features of corresponding classification algorithm. The optimal RF/SVM/PART classifier was constructed with its optimal features. Table 2 and Figure 3D summarizes the detailed performance of these three optimal classifiers. Evidently, the optimal SVM classifier was best, whereas other two optimal classifiers were almost at the same level.

Based on the above IFS results, we can select a best classifier on each dataset, which can be an efficient tool for identification of AD samples based on a certain brain structure.

On each brain structure, the optimal PART classifier always provided the lowest performance. However, it can provide more clues than other two optimal classifiers to uncover different methylation patterns within a certain brain structure of AD patients and control samples. By employing the optimal features used in the optimal PART classifier, PART learned rules based on all AD and control samples. These rules are listed in Table 3. Two rules were obtained for CRB brain structure. For other three brain structures, four, two and two, respectively, rules were accessed. These rules clearly displayed the different patterns on methylation levels between AD and control samples, which were helpful to improve our understanding on AD.

Cerebellum mainly contributes to learning and the establishment of theories and computational models. Given that a reduction in learning capacity is a typical symptom of AD (Sultana et al., 2006), identifying the methylation alterations during AD pathogenesis in the CRB is necessary. In our prediction list, the top methylation locus was cg07094785, which targets HNRNPUL2. Although only a few studies directly examined the function of this gene, it reportedly participates in AD pathogenesis, for example, by regulating the accumulation of tau protein in brain (Vanderweyde et al., 2016). Given that CRB tissues are highly specific, this gene was found to effectively function during nerve system degeneration due to age in multiple animal models, including mice and rats (Jasien et al., 2014; Kalinin et al., 2017). The other top-ranked locus was cg08105590, which targets FAM38A. This locus can also be associated with AD pathogenesis in the CRB at the methylation level. This gene might be correlated with AD by regulating Piezo1 channels (Velasco-Estevez et al., 2018), which has been shown to have a tight correlation with AD initiation and progression. Given that such correlations were validated using cerebellar cells (Velasco-Estevez et al., 2018), we can reasonably speculate that they can be observed in the CRB in vivo.

The first parameter for quantitative rules is cg15926585, which targets the intermediate region of COMT and TXNRD2. These two genes are both highly negatively correlated with the abnormal behavioral symptoms of AD (Borroni et al., 2006; Jin et al., 2017). Methylation of this region may inhibit the expressions of these genes and further trigger AD initiation and progression. Two other parameters that target HNRNPUL2 and FAM38A are functionally correlated with AD. Hypermethylation of these genes has also been demonstrated to participate in the development of diseases associated with nerve disorders (Pathak et al., 2019; Gürkan et al., 2020) but not with AD. Therefore, these quantitative rules in the CRB were explained according to publications above, demonstrating the effectiveness of our analysis.

Entorhinal cortex is a region that controls memory, navigation, and time perception (Goyal et al., 2018; Montchal et al., 2019). ERC is reportedly correlated with AD initiation and progression (Marzi et al., 2018; Grubman et al., 2019; Petrache et al., 2019). In our predicted feature list, the first locus is cg05810363, which targets RHBDF2. This gene is hypermethylated in HIPPO cells during AD pathogenesis (De Jager et al., 2014). Similar methylation patterns were recently identified in the ERC via interacting with H3K27AC (Marzi et al., 2017; Marzi et al., 2018). The methylation status of another locus, namely, cg08462501, which interacts with TBC1D1, is predicted to be downregulated during AD pathogenesis. Given that the ERC is correlated with a specific biomarker called TDP-43, this region could be hypomethylated during AD pathogenesis (Sun et al., 2017). Therefore, TBC1D1 may also present methylation status in the ERC similar to that in HIPPO.

Three quantitative rules, including the two features discussed in this section, i.e., the hypermethylated gene RHBDF2 and the hypomethylated gene TBC1D1, may contribute to AD pathogenesis. This contention was supported by previous reports.

Dorsolateral prefrontal cortex mainly controls the decision-making and working memory functions of the human brain. Given that people with AD show a decrease in the ability to render judgment (Heekeren et al., 2006) and a decline in working memory at late stages (Mars and Grol, 2007), we can reasonably regard this region as a potential target to identify methylation effects associated with AD pathogenesis. A locus predicted by our computational analysis is cg14622549, which targets the gene body region of EP400. As early as 2013, EP400 has already been regarded as a potential biomarker for detecting AD at the transcriptomics level (Desire et al., 2013; Trifonova et al., 2022). Methylation of EP400 is reportedly functionally correlated with autism, another disease of the nervous system that affects DLPFC. This supposition suggests the potential pathological effects of EP400 and supports our contention that methylation affects specific brain structures. The next locus is cg21097354, which targets the intermediate region of TTYH3 and AMZ1. AMZ1 has been shown to be a potential regulator for nerve system, especially in the elderly (Swingler et al., 2009; Panossian et al., 2018). Given that AD is a typical disease of the nervous system that is associated with aging, we can reasonably predict the potential relationships between DLPFC and AD. On the basis of a monozygotic twin model, a study suggested that the demethylation of AMZ1 in DLPFC is associated with dyskinesia, another potential symptom of AD (Soerensen et al., 2019). This report further confirmed the pathogenesis and downstream pathogenic regions of this methylation alteration.

The two parameters in the only classification rule were the two aforementioned loci. Their demethylation would aid in the identification of patients with AD according to this rule.

Numerous methylation loci can help in the identification of specific methylation alterations associated with AD in HIPPO, a region that regulates memory and navigation. The first locus is cg08056778, which targets the 3’UTR region of TBC1D1 and may further affect the expression of this gene in HIPPO. TBC1D1 is correlated with cerebral palsy (Mohandas et al., 2018). Moreover, TBC1D1 may be precisely regulated by effective microRNAs, thereby contributing to the development and aging of the central nervous system (Sephton et al., 2012). An in vivo study based on the model organism Caenorhabditis elegans confirmed that TBC1D1 in HIPPO may be functionally correlated with aging-related diseases (Escoubas-GüNey, 2018). Although this study did not cover AD, it nevertheless confirmed the tissue specificity of this gene. Thus, we can reasonably argue that pathological methylation alterations in HIPPO may affect AD pathogenesis, thereby validating our prediction. Similarly, SDK1, ANK1, and CHRNB4, which are targeted by cg04680535, cg05066959, and cg17179314, respectively, also have either a certain gene expression level regulated by a specific methylation status (De Jager et al., 2014; Kanno et al., 2014; Yan et al., 2019) or an epigenomic change during AD pathogenesis (Li et al., 2010; De Jager et al., 2014).

Two parameters associated with the extracted quantitative rules were predicted to aid in the identification of patients with AD. The first parameter is RHBDF2 (cg13076843), which has already been confirmed to be associated with AD (De Jager et al., 2014). The other parameter, that is, cg06639320, was also screened to be hypermethylated in patients with AD. The targeting of FHL2 by this locus is reportedly correlated with AD progression (Bennett et al., 2017; Altuna et al., 2019).